Vehicle detection using panchromatic high-resolution satellite images as a support for urban planning. Case study of Prague's centre

Thematic paper: Earth Observation for Smart City and Smart Region Vehicle detection using panchromatic high-resolution satellite images as a support for urban planning. Case study of Prague’s centre Peter Golej1— Jiri Horak1— Pavel Kukuliac1— Lucie Orlikova1 1Department of Geoinformatics, Faculty of Mining and Geology, VSB-Technical University of Ostrava peter[email protected] Abstract The optical sensors on satellites nowadays provide images covering large areas with a resolution better than 1 meter and with a frequency of more than once a week. This opens up new opportunities to utilize satellite-based information such as periodic monitoring of transport flows and parked vehicles for better transport, urban planning and decision making. Current vehicle detection methods face issues in selection of training data, utilization of augmented data, multivariate classification or complexity of the hardware. The pilot area is located in Prague in the surroundings of the Old Town Square. The WorldView3 panchromatic image with the best available spatial resolution was processed in ENVI, CATALYST Pro and ArcGIS Pro using SVM, KNN, PCA, RT and Faster R-CNN methods. Vehicle detection was relatively successful, above all in open public places with neither shade nor vegetation. The best overall performance was provided by SVM in ENVI, for which the achieved F1 score was 74%. The PCA method provided the worst results with an F1 score of 33%. The other methods achieved F1 scores ranging from 61 to 68%. Although vehicle detection using artificial intelligence on panchromatic images is more challenging than on multispectral images, it shows promising results. The following findings contribute to better design of object-based detection of vehicles in an urban environment and applications of data augmentation. Highlights for public administration, management and planning: • Detection of vehicle occurrence by geoinformation technologies provides essential data for planning transport infrastructure. • Classic machine learning methods achieve similar results to advanced CNN for detection of vehicles using a panchromatic image. • ArcGIS API for Python provides a suitable user environment and libraries for image processing using artificial intelligence • For better CNN results, a larger amount of training data and data augmentation are recommended. Keywords Vehicle detection, WorldView, Segmentation, Supervised classification, Neural network, Faster R-CNN Received: 28 September 2022 Received in revised form: 23 November 2022 Accepted: 02 December 2022 The paper was originally presented at the “GIS Ostrava 2022 Earth Observation for Smart City and Smart Region“ conference held on-line in March, 2022 (https://gisak.vsb.cz/gisostrava/index.php). Selected presentations from the conference were significantly extended and are now published in this volume as thematic papers exploring various topics related to usage of Earth Observation in smart city and smart region applications. 1 Introduction The rapidly growing number of vehicles causes various traffic problems (Tan et al. 2020), which are still difficult to measure and predict (Stuparu et al. 2020). Besides tremendous emissions caused by traffic in air, land, and water, there is a gigantic loss of time and money every day in the world due to vehicle traffic congestion (Keler et al. 2017; Stuparu et al. 2020;Tan et al. 2020). Tracked 108 © Jan Evangelista Purkyně University in Ústí nad Labem GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 Available online at content.sciendo.com movement of objects is nowadays widely available and used for various applications in our society (Keler et al. 2017). Such monitoring can improve the prediction of short-term and long-term traffic situations based on better understanding of traffic congestion propagation in time and space. Assessment of vehicle occurrence can be also utilized for planning transport infrastructure, fuel demand assessments, walkability studies, evaluation of local traffic density, assessment of emissions, noise, dust, or microclimate and other parts of local environmental impact assessment (Seenouvong et al. 2016). Vehicle monitoring is essential for some financial instruments (e.g., toll tax, parking fees). The remote sensing approach can be used to localize parked vehicles in a city (Seo & Urmson 2009;Zambanini et al. 2020). Current parking monitoring usually employs surveillance cameras and sensors which enable full coverage of large areas (Stuparu et al. 2020). Remote sensing powered by deep learning technologies and appropriate financial instruments could solve the problem of how to effectively limit short/long-term parking (Hong et al. 2022). Special applications can include detection of traffic incidents (Phiri & Morgenroth 2017) or terrorist monitoring (Husain et al. 2020). Satellite imagery provides snapshots in time that cover a relatively vast area. Compared with ground sensors, satellite remote sensing is more convenient and economic for obtaining urban traffic information (Tan et al. 2020). The progress of advanced processing techniques helps to solve the problem of big data, for which traditional techniques are not enough. The deep learning revolution has opened up an entirely novel frontier. Modern machine learning has boosted techniques such as the detection and automatic counting of objects, semantic segmentation and image classification (Merodio Gómez et al. 2021). The accuracy of vehicle detection from highresolution images can currently reach more than 90%. It is conditioned by good pre-processing, including feature extraction such as Hough transformation and other methods of edge detection (Eslami & Faez 2010). Current research prefers to avoid direct extraction of features and focuses instead on advanced ML methods such as RetinaNet Architecture, Faster R-CNN or YOLO (Bin Zuraimi & Kamaru Zaman 2021;Ghosh 2021;Ma et al. 2019; Stuparu et al. 2020;Tan et al. 2020), which can reach more than 94%. The accuracy is not always so high; results depend on many factors, such as environment. In urban areas, the accuracy is usually lower, even when using the same method (e.g. 65% for Faster R-CNN in Zambanini et al. 2020). Also, in the case of lower spatial resolution, results are worse (e.g. accuracy less than 70% in Chen et al. 2021). Object-oriented classification of satellite imagery still represents a challenge Zhang et al. 2020a, partly due to the wide spectrum of required tasks such as object detection, object tracking, image segmentation and remote sensing image interpretation. Object-based methods dominate among classification techniques of remote sensing images, however, pixel-based methods or sub-pixel-based methods are also still used (Li et al. 2014). Object-based image analysis (OBIA) consists of two steps: image segmentation and object classification (Li et al. 2014;Vecer et al. 2021). The choice of segmentation parameters depends on the specialist performing the required task and the selection is usually subjective and arbitrary. Inappropriate selection of segmentation parameters can lead to undersegmentation or over-segmentation, negatively influencing the final classification (Golej et al. 2021; Zanotta et al. 2018). Object-based image analysis classification methods typically include SVM (Support Vector Machine), KNN (K-nearest neighbor classifier) or RT (Random Trees) (Abburu & Babu Golla 2015;Hossain & Chen 2019;Phiri & Morgenroth 2017;Yekkehkhany et al. 2014). Object detection is one of the main research areas of computer vision. In the past decade, numerous deep learning based object detection methods for images in remote sensing have been designed for detection of specific objects (Hou et al. 2020;Koga et al. 2018;Xiao et al. 2020;Zhao et al. 2017). The traditional vehicle detection algorithms such as the Gaussian mixed model (GMM, Stauffer and Grimson 1999), give overall satisfactory results, however, they may not work properly when illumination changes occur or due to background cluttering (Maity et al. 2021). The deep learning methods offer an inherent extraction capability feature which makes them much more acceptable to scholars compared to traditional methods as it minimizes the errors occurring in classification tasks (Maity et al. 2021). Convolutional neural networks (CNN) are designed to learn the spatial features, such as edges, corners, or textures, that best describe the target class. The core for learning these features are manifold and successive transformations of the input data (convolutions) on different spatial scales (Kattenborn et al. 2021). The main advantage of CNN is that it automatically detects significant spatial features without any human supervision (Alzubaidi et al. 2021). As CNNs are designed to artificially replicate the functional capabilities of a human cognitive system, they perform better in var- © Jan Evangelista Purkyně University in Ústí nad Labem 109 Available online at content.sciendo.com GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 ious computer vision tasks compared to the traditional methods (Benjdira et al. 2019;Maity et al. 2021). Currently, the most popular target detection algorithms include RCNN, Fast RCNN, and Faster RCNN (Hou et al. 2020;Rawat 2019;Zhang et al. 2020b). Faster R-CNN is the state-of-the-art algorithm used for generic object detection and it has been successfully adapted to account for many recognition problems. Faster R-CNN has two predecessors: RCNN and Fast R-CNN. R-CNN (Region-based Convolutional Network) is one of the primary deep neural networks which are designed to perform object detection (Maity et al. 2021). A selection search algorithm extracts around 2000 region proposals which might contain objects (Girshick et al. 2014; Maity et al. 2021). Each of these region proposals or Regions of Interest (RoIs) is processed through a convolutional neural network (CNN) to obtain feature maps. The feature maps are then classified using the SVM model and a bounding box regressor is used to obtain bounding boxes (Girshick 2015;Girshick et al. 2014;Maity et al. 2021). However, this method was criticized for high complexity and the resulting time demands which limits its real applications (Girshick 2015;Maity et al. 2021). Fast R-CNN was developed by (Girshick 2015) and was intended to accelerate processing. In this model, instead of feeding each of the 2000 regions to separate CNNs, the whole image is fed to a single CNN. An RoI pooling layer is used to extract and resize the feature maps of all the region proposals to the same size (Maity et al. 2021). This is then passed on to fully connected layers consist of two branches: a softmax classifier to give probabilities for each class, and a bounding box regressor for precise bounding box coordinates (Girshick 2015). The RoI pooling layer speeds up the object detection of Fast R-CNN compared to R-CNN. However, the problem of inaccurate region proposals still exists in Fast R-CNN due to the non-learning capability of the selective search algorithm (Maity et al. 2021). Faster R-CNN is divided into two modules: the Region Proposal Network (RPN) and a Fast R-CNN detector. Faster R-CNN improves the object detection architecture by replacing the selection search algorithm in Fast R-CNN with a Region Proposal Network (RPN) (Benjdira et al. 2019;Li et al. 2020). RPN is a fully convolutional network used to generate region proposals, but also can simultaneously propose object bounds and object scores at each position (Li et al. 2020;Maity et al. 2021). The rest of the model architecture remains the same as Fast R-CNN. This design increases the speed of detection and brings it closer to real time (Benjdira et al. 2019; Maity et al. 2021). The system overview of Faster RCNN is given in Fig. 1. In order to detect vehicles more accurately and provide accurate traffic information effectively and timely, this study uses object-oriented classifications and also deep learning algorithm to detect vehicles in panchromatic images from high-resolution satellite remote sensing technology. The aim of the paper is to demonstrate differences and issues in vehicle detection using a panchromatic satellite image of the old city centre, and to compare usage of the selected traditional machine learning methods with the advanced Faster R-CNN method. The paper is organized as follows: after the introduction, the study area and data sources are described. Further, processing in selected SW tools is explained focusing on the explanation of hyperparameters used for processing settings. Next, the obtained results and calculated quality parameters describing the success of vehicle detection methods are provided. Finally, results and various limitations of these methods are discussed. Fig. 1 System overview of Faster-RCNN (Maity et al. 2021) 2 Methods 2.1 Study area and data sources The study area (Fig. 2) is situated in the centre of Prague around the Old Town Square, with an approx. size of 460 x 300 m. This area represents a historical centre with both narrow and curved old streets, newly built boulevards and various squares, including vehicles on streets as well as in parking lots. 110 © Jan Evangelista Purkyně University in Ústí nad Labem GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 Available online at content.sciendo.com Among various high resolution satellite imagery, one scene of WorldView3 was obtained. Spatial resolution of the panchromatic and multispectral WorldView3 images is 0.3 m and 1.6 m, respectively. The full scene (July 23, 2019, morning) covers 25 km2from which a small subset of the panchromatic image was selected to establish the study area. As appropriate auxiliary data, vector data were obtained for buildings in Prague (Prague Geoportal 2021). This data enabled the creation of a mask to improve the quality of detection. Assessment of the accuracy of vehicle detection requires knowing the ground truth. On-line manual vectorization of vehicles in the study area was performed by a specialist. Three types of vehicles were distinguished – dark, bright, and vehicles in shadow (Table 1); the total number of vehicles is 264. The average size of vectorized vehicles is 7.89 m2 which corresponds to the average size of the top 10 best-selling cars in the Czech Republic, which is 7.94 m2 (ePojisteni.cz 2022). As expected, the size of vehicles in shadow is smaller due to worse conditions for detection. Table 1 Ground truths of observed vehicles in the study area Count Average vehicle size (m2) Dark vehicles 114 7.89 Bright vehicles 104 8.25 Vehicles in shadow 46 7.17 3 Methodology First, the image was masked. The mask filtered out places where no vehicles or people were expected. Masking limits the detection of objects to the road network, and public spaces such as streets or squares, which provides better classification results. Three different software packages were used for vehicle detection, image segmentation, modelling and classifications – ENVI, CATALYST Pro, and ArcGIS Pro. The results from all three systems were compared and evaluated. 3.1 Segmentation and classification in ENVI ENVI is a software dedicated to image processing and analysis, integrated in Esri´s ArcGIS platform. ENVI version 5.6 was used in the research for this paper. ENVI offers two algorithms for segmentation settings. The first one is the edge detection algorithm which is suitable for clearly defined objects. The second is the intensity algorithm, which performs best for images with subtle gradients such as digital elevation models. ENVI also offers two algorithms for merge settings. In this study, the Full Lambda Schedule algorithm was used because it is more efficient in heterogeneous urban areas with a variable texture. It is necessary to set up the merge value appropriately to avoid oversegmentation or under-segmentation. The Texture Kernel Size can be specified as an odd number between 3 and 19. In our case, The value was set to 3 to assure minimal segment areas for such variable urban conditions. After image segmentation, training data were obtained. In total, eight training classes were created - road, shadow, vegetation, crosswalk, sidewalk, bright vehicles, dark vehicles and vehicles in shadow. The classes of crosswalk and sidewalk were added due to the similarity of their pixels to one of the three vehicle types. The second step of OBIA processing is image classification. ENVI provides the following classification methods: KNN, SVM and PCA. For all three methods, after preliminary testing, the threshold value was kept at the default value 5. Each segment was assigned to the class with the highest confidence value. Segments with class confidence values below the threshold value were left unclassified. In the case of the K-NN method, K (number of neighbors) was chosen to be 5. This is because an odd integer ranging from one to a value less than or equal to the total number of training classes for all classes needed to be set. SVM offers four kernel types. The Radial Basis function (ThurnhoferHemsi et al. 2020) was selected for this SVM classification. Further, spectral, textural, and spatial attributes for classification must be chosen. The best combination of attributes recommended by ENVI was selected for KNN and SVM, while for PCA all attributes were used. 3.2 Segmentation and classification in CATALYST Pro CATALYST Professional is a geospatial desktop suite specializing in Remote Sensing, Photogrammetry & Earth Observation Science for optical and SAR imagery. CATALYST PRO version 2222.0.3 was used in this study. Segmentation in CATALYST Pro requires setting three main segmentation parameters: scale (corresponding to the object size – the larger the scale, the larger the output object will be), shape (representing the weight of the shape for segmentation in balance with the colour weight) and com- © Jan Evangelista Purkyně University in Ústí nad Labem 111 Available online at content.sciendo.com GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 Fig. 2 Study area (Old Town Square in Prague, WorldView3, panchromatic image, resolution of pixel is 30x30 cm) pactness (denoting the compactness of the shape for segmentation). A higher compactness value generates more compact object boundaries, suitable for, i.e., crop fields. For the second step of OBIA, CATALYST Pro offers two classification methods: RT and SVM (again with four kernel types). Similarly to ENVI’s settings, the Radial Basis kernel was selected for SVM. CATALYST Pro feeds the segmentation process with many enumerated attributes organized into the following groups: statistical, geometric, vegetation indices, textural, and polarimetric. For vehicle detection, the following attributes were selected: the mean from the statistical group, and the compactness (unitless geometric measure of a shape, independent of scale and rotation, Hage and Hamade 2016), elongation (ratio between the major axis and the minor axis, Fernandes et al. 2014), circularity (Xu et al. 2017), and rectangularity (the ratio of the segment area to the area of its minimum bounding rectangle, Guindon et al. 2004) from the geometric group. 3.3 Convolutional neural networks Finally, vehicle detection using artificial intelligence was performed, specifically a convolutional neural network and its Faster R-CNN algorithm. The ArcGIS Pro environment using the ArcGIS API for Python was used to automate the detection process. Contrary to detections in ENVI or CATALYSTO Pro, the detection with the Faster R-CNN algorithm applied a mask of buildings at the final stage, after the training and testing phases. Testing of the use of a preliminary mask at the beginning of the classification process with Faster R-CNN failed probably due to insufficient training data. Also, all three types of vehicles were joined into one type. First, it was necessary to create an image that contained three bands (such as RGB), because neural network algorithms work better with such an image type (Kumar et al. 2018). The “composite bands tool” was used to create an artificial image with 3 bands from the panchromatic image. Next, it was 112 © Jan Evangelista Purkyně University in Ústí nad Labem GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 Available online at content.sciendo.com necessary to export the training data. The online manually vectorised data were used as training data (Table 1). In the training phase, it was necessary to set the following parameters: the tile size (the size of the image chips) was set to 128. A smaller size did not improve the results, and a larger size was limited by the current hardware performance. A stride size (the distance to move in the X and Y when creating the next image chip; when stride is equal to half of the tile size, there will be 50% overlap) was set to 64. The metadata format of training data was set to PASCAL Visual Object Classes in .xml format, which is suitable for the further processing. The maximum number of training epochs was set to 50. Faster R-CNN was selected with the maximal available batch size 32. After optimization, the backbone model was set to ResNet-50 providing the best results. The percentage of training cases for validation was set to 30%. In the testing phase, the probability threshold for the detection of objects (vehicles) was set to 0.65, and the batch size to 32. After the training and testing phase, the mask of buildings was used as a final step. It helped to eliminate a number of False Positive vehicles detected on rooves. 4 Results All methods were applied to detect vehicles in the pilot area. The number of detected vehicles varied between 142 (Faster R-CNN) and 596 (PCAe) (Table 2). As it is shown in the table below, the number of vehicles from PCA-e is overestimated (533 vehicles detected). Processing speed is relatively high, especially in the case of ENVI and CATALYST Pro. Detection using neural networks compared to OBIA is more time-consuming. However, the most time-consuming was the vectorization of vehicles by a specialist and the creation of training samples for classification. This was mainly due to the differentiation of vehicle types, differentiation of vehicles from other classes such as shadow, or the selection of representative segments for training classes. Table 2 Number of detected vehicles Vehicles KNN-e SVM-e PCA-e SVM-c RT-c Faster R-CNN Vectorization Bright 119 88 55 54 97 - 104 Dark 217 172 533 88 111 - 114 In shadow 20 20 8 11 31 - 46 Total 356 280 596 153 239 142 264 KNN-e, SVM-e, and PCA-e denote methods implemented in ENVI; SVM-c and RT-c methods implemented in CATALYST Pro Results were compared with the outputs of online vectorization which is considered as a ground truth (Table 3). It enables the evaluation of the number of vehicles successfully detected by the given method (True Positives – TP), the number of nonvehicle objects that are falsely detected as vehicles, (False Positives – FP) and the number of vehicles which the method did not recognize as vehicles but actually were vehicles (False Negatives – FN) (Benjdira et al. 2019). Table 3 Evaluation parameters of using methods Method True Positives False Positives False Negatives KNN-e 193 163 71 SVM-e 200 80 64 PCA-e 143 453 121 SVM-c 128 25 136 RT-c 169 70 95 Faster R-CNN 127 15 137 KNN-e, SVM-e, and PCA-e denote methods implemented in ENVI; SVM-c and RT-c methods implemented in CATALYST Pro The behaviour of each method is exemplified in the following figures showing the area of Kaprova street. The main issue in classification seems to be an occurrence of mixed pixels. Especially, the KNN method in ENVI was unable to detect many vehicles (Fig. 3a left). Better results can be seen in the case of SVM, but still classification deteriorates for vehicles with a high frequency of mixed pixels (Fig. 3a right). The main problem could be the similarity of the mixed pixels of the vehicle and the mixed pixels of the sidewalks, or the similarity of the mixed pixels of sidewalk and crosswalk, which causes worse vehicle detection results. Similar issues with “mixed” pixels are found in the results of the RT method in CATALYST Pro (Fig. 3b left). A further issue is the detection of vehicles in shadows (both tree and building shadows). © Jan Evangelista Purkyně University in Ústí nad Labem 113 Available online at content.sciendo.com GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 The SVM method in CATALYST Pro (Fig. 3b right) could not detect a vehicle in the shadow. The PCA method in ENVI in this situation achieved the worst result (Fig. 3c left). This method failed to detect vehicles in shadow, bright vehicles, and vehicles that contain “mixed” pixels. The PCA method also was not able to detect bright vehicles on Kaprova street. Conversely, Faster RCNN was able to detect vehicles in shadow, and dark as well as bright vehicles, so it detected all the vehicles in the selected area of Kaprova street (Fig. 3c right). False positive vehicles were detected in ENVI and CATALYST Pro SW mainly in shadow areas and in road areas, due to pixel similarity to darker vehicles (Figure 3d). An issue in the usage of the Faster R-CNN method is false classification of some parts of rooves as vehicles. This is due to the fact that the building mask was not used before the detection (Fig. 3e left). That is why masking was applied after the vehicle detection and the number of FP vehicles was markedly reduced. Resting FP vehicles were mainly detected on roads or in empty parking lots containing “mixed” pixels – usually the result of horizontal traffic signs (Fig. 3e right). After processing the panchromatic image, three quality indicators were enumerated to evaluate the accuracy of the vehicle detection - Precision, Recall and F1 score (Table 4) (Benjdira et al. 2019). In order to calculate these indicators, it was necessary to determine TP, FP and FN (Table 3). The best true detection of vehicles was obtained using the SVM method in ENVI. Oppositely, the Faster R-CNN algorithm and SVM method in CATALYST Pro did not recognize the majority of vehicles, so the resulting FN value was the highest of all. The highest FP indication was provided by the PCA method in ENVI. Faster R-CNN detected the least FP vehicles after using the mask. Table 4 Evaluation parameters of using methods Method Precision Recall F1 score KNN-e (%) 54.21 73.11 62.26 SVM-e (%) 71.42 75.76 73.53 PCA-e (%) 23.99 54.17 33.26 SVM-c (%) 83.66 48.48 61.39 RT-c (%) 70.71 64.02 67.19 Faster R-CNN (%) 89.44 48.11 62.56 KNN-e, SVM-e, and PCA-e denote methods implemented in ENVI; SVM-c and RT-c methods implemented in CATALYST Pro As can be seen from Table 3, evaluation parameters show that, Faster R-CNN and SVM in CATALYST Pro have a high precision rate (89.44% for Faster RCNN and 83.66 for SVM). This high value indicates that when they classify an object as a vehicle, it is very highly probable that this object is in fact a vehicle. So, the ability of these algorithms and methods to detect real vehicles is quite high. But, when comparing the recall, it is apparent that the best results are achieved by SVM in ENVI and then KNN in ENVI (75.76% and 73.11). The recall measures the ability of the algorithm to detect all the instances of vehicles in an image. Faster R-CNN and SVM in CATALYST Pro achieve the worst results (48.11% and 48.48%). SVM and KNN in ENVI are more capable of extracting almost all instances of vehicles, while Faster R-CNN and SVM in CATALYST Pro miss the majority of vehicles. Considering F1 score, which is a harmonic mean of the precision and recall that gives an assessment of the algorithms’ effectiveness, it is possible to summarize that the best results were achieved by SVM in ENVI. Of course, the suitability depends on which type of errors the end-user prefers to eliminate. 5 Discussion The Faster R-CNN algorithm detected the least number of vehicles (143). The PCA method achieved the largest number of detected vehicles (596), but in many cases they were FPs. The SVM-c method detected a similar number of vehicles to Faster RCNN. Approximately the same number of vehicles compared to vectorization was achieved by the RT-c (239) and SVM-e (280) methods. The KNN-e method detected a total of 356 vehicles. High FP detection was usually caused by mixed pixels, mainly on roads (especially caused by horizontal traffic signs) but also on crosswalks or sidewalks with bright vehicles or vehicles in the shadow. The PCA method has the largest number of FPs (almost double the total number of vehicles), which confirms the inappropriateness of using this method for classification. The least FPs were detected by Faster R-CNN. The main problem is segmentation, where some vehicles are lost due to their low contrast. Eikvil et al. (2009) presented an automatic approach consisting of a segmentation step followed by object classification to detect vehicles in high-resolution satellite images with 0.6m resolution. Approximately 80% of the vehicles obtained by vectorization were detected. Similar results were also obtained in our case, mainly using the KNN and SVM methods in ENVI. Eslami and Faez (2010) presented a framework to detect vehicles from high-resolution panchromatic images (0.6m) in non-urban areas. 114 © Jan Evangelista Purkyně University in Ústí nad Labem GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 Available online at content.sciendo.com Fig. 3 Detection of vehicles by KNN (left) and SVM (right) methods in ENVI (resolution of pixel is 30x30 cm); b: Detection of vehicles by RT (left) and SVM (right) methods in in CATALYST Pro (resolution of pixel is 30x30 cm); c: Detection of vehicles by PCA method in ENVI (left) and using Faster R-CNN (right)(resolution of pixel is 30x30 cm); d: Incorrectly detected vehicles using KNN method in ENVI (image on the left - shadow area; image on the right - road area. Resolution of pixel is 30x30 cm); e: Incorrectly detected vehicles using Faster R-CNN algorithm (resolution of pixel is 30x30 cm) © Jan Evangelista Purkyně University in Ústí nad Labem 115 Available online at content.sciendo.com GeoScape 16(2) — 2022: 108—119 doi: 10.2478/geosc-2022-0009 They used feature extraction and image processing techniques like Hough transformation, gradient, and thresholding operation. They detected about 94% of total vehicles. With the development of deep learning and artificial neural networks, selected deep learning methods are also being used in high-resolution remote sensing based vehicle detection in recent years (Ma et al. 2019). As stated by Stuparu et al. (2020), an F1 score of 94.48% was achieved when using the RetinaNet Architecture for object detection and the Cars Overhead With Context dataset (COWC) with a 0.15m resolution. Chen et al. (2021) utilized multitemporal Planet satellite image data (3 m) to develop a vehicle detection method. This vehicle detection algorithm achieved an F1 score of 68% due to low spatial resolution. Zambanini et al. (2020) presented an approach to localize parked vehicles in a city, where a Faster R-CNN detector was trained and applied to stereo satellite images (WorldView 3) to discriminate static from moving vehicles. This model achieved an average precision of 65% for parked vehicle detection. Arora et al. (2022) detected vehicles using Fast R-CNN technology for efficient detection on the road, in both day and night mode. They performed automatic detection of moving vehicles on the road using realtime video of front-side vehicles. Precision was 90%. Ghosh (2021) proposed a method for onroad vehicle detection in varying weather conditions using Faster R-CNN. Three different public datasets, DAWN, CDNet 2014 and LISA, were used for detection. Precision reached the level of 89.48% for DAWN, 91.20% for CDNet 2014 and 95.16% for LISA. Tahir et al. (2022) presented an approach for detection of aircraft objects from Google Earth images. They used Faster R-CNN, YOLO and SSD with accuracy 95.31%, 94.20% and 84.61%, respectively. As Rawat, 2019 describes, Faster R-CNN excels in its task of object detection. He recommends enhancing the model performance by generating higher quality images using data augmentation. To detect various objects (ships, planes, storage, bridges and harbours) in a panchromatic image, Hou et al. (2020) used Faster R-CNN, SSD and YOLOv2 algorithms where the results were very good mainly for ship, storage and bridge detection. The main detection problem was the detection of planes because they have extremely small sizes and plenty of intraclass, which is similar to vehicles detection. The procedure for vehicle detection using Faster R-CNN proposed in this paper uses very highresolution satellite images, namely panchromatic images with a resolution of 0.3m. The Faster R-CNN algorithm used achieves an F1 score of 62.56%. Stuparu et al., 2020 used the RetinaNet architecture and achieved an F1 score of 94%. This may be due to having more data and better data resolution (0.15m). Although Chen et al. (2021) achieved a better F1 score (68%) with lower resolution satellite images, this value may be affected by more available training data. Precision in our case reached almost 90%. Arora et al. (2022) detected vehicles using images created from videos, where precision was also around 90%. Ghosh (2021) achieved precision values ranging from 89% to 95% on a more robust dataset. Tahir et al. (2022) detected aircraft objects using Google Earth images with better resolution than our satellite images. The Faster R-CNN algorithm achieved an F1 score of 95.31%. In each of these cases, a larger number of satellite images or images were provided (in most cases with better resolution), which resulted in a larger amount of training data, therefore, the results of the individual algorithms were better. As Rawat (2019) notes, it will be necessary to carry out data augmentation in the future for a larger amount of training data, which will help us achieve better results. The current hardware limitation (Intel Core i7-8700 CPU 3.20GHz; NVIDIA GEFORCE GTX 1070, and 16 GB RAM) does not enable full optimization of the advanced machine learning process. CUDA (Gavali & Banu 2019) was installed to improve the calculation. CUDA divides more complex calculations into simpler ones and distributes them in parallel among threads in a Graphic Processing Unit (GPU). The improvement in GPU performance over CPU performance is usually 10-20:1 (Alzubaidi et al. 2021) The recommendation of the best method for vehicle detection depends on the user’s preferences. If the goal is to detect as many vehicles as possible, the Faster R-CNN algorithm (89%) seems to be the best choice, followed by SVM in CATALYST Pro (84%). If the objective is to minimize missed vehicles, scholars should prefer SVM in ENVI (76%) followed by KNN in ENVI (73%). According to the F1 score, the SVM method in ENVI appears to be the best (74%), followed by the RT method in CATALYST Pro (67%). Object-oriented classification has its limitations. Advanced image processing methods are available in relatively expensive commercial software or in free open-source software requiring knowledge of programming languages. Another limitation can be the loss of target objects during segmentation, e.g., of dark vehicles in the shadows. An important step that can improve segmentation is the creation of a mask to filter out areas where vehicles cannot be located, however, such data is not always 116 © Jan Evangelista Purkyně University in Ústí nad Labem